Extending a Scheduler to Better Support Sharded Services

Sharded services are incredibly common in modern computing, but schedulers don't usually consider them when making scheduling decisions, preferring to take a task based view of service health. This disparity between models can lead to ugly workarounds, which can in some case jeopardize the reliability of a service. This talk describes a project that extends a task based scheduler to take shards into account when scheduling. This project has sped up updates and reduced downtime for many services across Facebook, and has a surprisingly diverse set of uses.